We present some use cases to illustrate the strength of using the combination of machine learning and operations research. We only illustrate the first combination but there are four ways to combine machine learning and operations research .
Quite often, we improve processes by 10, 20 or even 40% when using this combination of machine learning and operations research. You can discover some of our case studies .
You want to send a fleet of UAVs to do some collective surveillance of an ongoing fire. You want to collect as much relevant data as possible without taking a high risk with your vehicles. Also you want to minimise the number of vehicles sent while at the same time maximize the relevant areas covered.
1. You use machine learning for scene classification to assess where the fire is, what the dangers of the fire are and maybe even discover navigable paths for the UAVs. 2. You use operation research to first navigate to the most relevant parts of the areas while minimizing the fleet and mitigating the risks.
Machine learning is able to extract representative features from the raw measurements provided by sensors on board of the UAVs. It can assess the presence and dangers of fire.
Operations research will optimize and plan the search by dividing the areas among the UAVs and mitigate the risk of loosing some vehicles by first accessing less dangerous parts of the areas and then only the most dangerous parts if needed.
You are an automobile manufacturer and you want to implement this incredible Self-Driving Car Technology.
1. With machine learning you can improve the inputs from your sensors and detect the road conditions, the neighboorhood of the car, your driving conditions, etc. 2. Given these improved inputs, you want to optimize your driving and above all the safety of your car.
Machine learning can help you detect and better "understand" your surroundings. With computer vision, you can detect objects, with LIDAR you can have a 3D representation of your surroundings.
Operation research can optimize the driving for safety, energy, time, etc. For this specific example, we will even go further and mention that for safety reasons only, you absolutely should add some operation research to help machine learning perform better and safer. For instance, deep learning is not able to detect and/or recognize all objects in real time in a safe manner. Adversarial examples pop up all the time and just for that alone, you need to add a layer of operations research.
Your company builds planes and you want to improve the overall process: you want to build planes faster and cheaper.
1. With machine learning you can follow and predict what teams are/should do. 2. With this data, operations research will optimize the overall process while also ensuring that the numerous specifications are met.
Once a first planning is in place with a mixed top-down and bottom-up approach, machine learning can help you follow in real time any deviation from your teams and subcontracters and if your specifications are respected.
Operation research in turn will optimize the workflow between your teams and subcontractors in real time but also optimize your plannification in the short, medium and long terms.
Your are a port and you want to optimize the loading and unloading of the myriads of containers transiting through your port.
1. With machine learning you can follow and predict the actions of all involved actors: boats, trucks, cranes, customers, inspectors, customs, etc. 2. Operations research will then optimize the overall loading and unloading processes.
Machine learning can predict the arrival times, the congestion, the service time, the loading and unloading times, etc.
Operation research can optimize the loading and unloading operations by placing the containers in the right order, moving the crane optimally, ordering/scheduling the incoming and outcoming trucks, etc.
You are a train company and you own kilometers of tracks that you need to repair or replace from time to time. You want do to that for a minimum cost.
1. You take pictures of the tracks and with Machine Learning you detect possible defects. 2. You define (optimal) repair/replacement policies with Operations Research.
You put cameras under your wagons and take pictures of your tracks in rapid succession. Deep learning and computer vision techniques let you stitch your pictures of railways and not only detect your defects but also recognize their importance. For instance, you can tell if a defect needs to be addressed in 3 months or in two years.
Once you have your network/graph of defects, you need to act. The idea is that it might be worth reparing a track that does not need immediate repair: it is more efficient to let the team already onsite do some maintenance instead of sending it back a few months later. Only Operations Research can help you devise such repair/replacement policies.
You are a delivery company and you want to deliver some goods to your customers. You want to do this in a timely and efficient manner for a minimum cost.
1. With machine learning you can predict for each of your driver how long they would drive to visit and serve your customers. 2. You then can optimize the routes.
Machine learning will give you better estimates of the travel and serve times taking into account not only the driver but also the hour of the day, the trafic, the weather, the period of the year, etc.
Operation research optimizes your routes while at the same time respecting some constraints like time windows, work hours, priorities, etc.
You keep your goods in a warehouse and want to retrieve your goods as quickly as possible while minimizing the transportation risks and any possible mistakes in the retrieval process.
1. With machine learning you will not only predict how long it takes to retrieve good A from shelf B but also what is likelihood of a demand for a given good on a given day. 2. Given these time and demand estimations, use operations research to optimize where you place your goods but also how you retrieve them.
Machine learning can better predict the real time and energy needed to take a good A from a shelf B as well as realistically assess some of the risks involved. Machine learning will also help you forecast the demand for a given day, thus you will be able to pro-activaly rearrange some goods for a better, quicker and/or less error prone process.
Operation research can optimize where to locate your goods, if you need to keep them at one place or divide them into several places, how you can retrieve these goods. Operation research can also minimize the risks based on the risk assessment from machine learning predictions.
(Special dedication for MTLab). You want to launch a startup to work on multi-modal trips. Because you know that people want more than just going from point A to point B to point C, you add some fancy points of interests in the mix. You still want to propose realistic journeys and don't want to overflow your customers with too many possibilities and customize as much as possible their trips.
1. With machine learning you learn what the points of interest for your customers are and what they most likely will love to do. 2. Operations research lets you optimize the trajectories while taking into accounts the different possibilities to move and points of interest along the way.
Machine learning helps you detect the trends, what touristic monuments your customers are most likely to visit or not and at what time of the year, how much time they will spend to visit them, etc. You can also analyze customer reviews with NLP to understand what they like or not.
Operation research optimizes your trajectories and what to visit or not along the way. You can optimize the travel times, the costs, the joy of discovering new things, etc or a combination of all these objectives together.